Non-ASCII Python identifiers and reflectivity [duplicate]

This question already has an answer here:

• Identifier normalization: Why is the micro sign converted into the Greek letter mu? 1 answer

I have learnt from PEP 3131 that non-ASCII identifiers were supported in Python, though it's not considered best practice.

However, I get this strange behaviour, where my 𝜏 identifier (U+1D70F) seems to be automatically converted to τ (U+03C4).

class Base(object):
    def __init__(self):
        self.𝜏 = 5 # defined with U+1D70F

a = Base()
print(a.𝜏)     # 5             # (U+1D70F)
print(a.τ)     # 5 as well     # (U+03C4) ? another way to access it?
d = a.__dict__ # {'τ':  5}     # (U+03C4) ? seems converted
print(d['τ'])  # 5             # (U+03C4) ? consistent with the conversion
print(d['𝜏'])  # KeyError: '𝜏' # (U+1D70F) ?! unexpected!

Is that expected behaviour? Why does this silent conversion occur? Does it have anything to see with NFKC normalization? I thought this was only for canonically ordering Unicode character sequences...